NBA net rating quadrant

The brief: a season-review blog post needs one chart of where all 30 teams finished on offense and defense, split by conference, at 1600 px wide, plus a square version for Instagram. This recipe builds it in plotnine, where each fix is one more layer: logos with geom_sdv_logos, per-conference averages with geom_mean_lines, and facets for the conferences. The ratings come from hoopR's ESPN team box scores through sportsdataverse.nba.

import tempfile
import warnings
from pathlib import Path

import polars as pl
import sportsdataverse.nba as nba
from IPython.display import Image
from PIL import Image as PILImage
from plotnine import (
    aes,
    element_blank,
    element_line,
    element_text,
    facet_wrap,
    geom_label,
    geom_line,
    geom_point,
    geom_text,
    ggplot,
    labs,
    scale_x_continuous,
    scale_y_reverse,
    theme,
    theme_minimal,
)

import sdvplot
from sdvplot.plotnine import geom_mean_lines, geom_sdv_logos

SEASON = 2026  # the 2025-26 season: NBA seasons are named by the year they end
LABEL = f"{SEASON - 1}-{SEASON % 100:02d}"
OUT = Path(tempfile.mkdtemp(prefix="sdvplot-recipe-"))  # where the exports go; use your own folder

1. Get the data

Offensive and defensive rating are points scored and allowed per 100 possessions. Possessions are estimated from the box score (FGA - OREB + TOV + 0.44 FTA) and averaged with the opponent's, so both teams in a game share one count.

box = nba.load_nba_team_boxscore(seasons=[SEASON]).filter(pl.col("season_type") == 2)  # regular season
possessions = (
    pl.col("field_goals_attempted")
    - pl.col("offensive_rebounds")
    + pl.col("total_turnovers")
    + 0.44 * pl.col("free_throws_attempted")
)
games = box.with_columns(poss=possessions)
opponents = games.select("game_id", opponent_team_id="team_id", opp_poss="poss")
assert games.schema["opponent_team_id"] == opponents.schema["opponent_team_id"]  # same dtype on both sides
games = games.join(opponents, on=["game_id", "opponent_team_id"]).with_columns(
    game_poss=(pl.col("poss") + pl.col("opp_poss")) / 2
)
ratings = (
    games.group_by("team_abbreviation")
    .agg(
        ortg=100 * pl.col("team_score").sum() / pl.col("game_poss").sum(),
        drtg=100 * pl.col("opponent_team_score").sum() / pl.col("game_poss").sum(),
    )
    .with_columns(net=pl.col("ortg") - pl.col("drtg"))
    .sort("team_abbreviation")  # group_by returns groups in any order; sort for a stable result
)
ratings.height

33

33 teams in a 30-team league. The extra three are the All-Star Game's teams, which ESPN files as regular-season games. resolve is the quickest way to find them: anything that is not an NBA team comes back None, with one warning that names it. The conference comes from the same resolved ids.

with warnings.catch_warnings(record=True) as caught:
    warnings.simplefilter("always")
    team_ids = sdvplot.resolve(ratings["team_abbreviation"].to_list(), "nba")
print(caught[0].message)

conferences = sdvplot.teams("nba").select("team_id", conference=pl.col("conference").str.replace("ern Conference", ""))
ratings = (
    ratings.with_columns(team_id=pl.Series(team_ids, dtype=pl.Utf8))
    .filter(pl.col("team_id").is_not_null())
    .join(conferences, on="team_id")
    .sort("net", descending=True)
)
ratings.head()

3 value(s) did not resolve to a nba team: 'STARS' (unknown), 'STRIPES' (unknown), 'WORLD' (unknown). Use sdvplot.suggest() for candidates, or strict=True to raise.

team_abbreviation	ortg	drtg	net	team_id	conference
OKC	115.988049	105.126054	10.861996	25	West
DET	114.414822	106.488601	7.926221	8	East
SA	116.576118	108.717581	7.858537	24	West
BOS	117.222701	109.368854	7.853846	2	East
NY	116.794332	110.394973	6.399359	18	East

2. The first draft

Points and abbreviations, plotnine's defaults.

(ggplot(ratings, aes("ortg", "drtg", label="team_abbreviation")) + geom_point() + geom_text(nudge_y=0.4, size=8))

Three problems: the labels overlap, a low defensive rating is good but sits at the bottom, and the two conferences (which only meet a third of the time) are mixed together.

3. Logos, one panel per conference, and average lines

geom_sdv_logos replaces the points (the team aesthetic takes the abbreviations as they are), facet_wrap splits the conferences, and geom_mean_lines draws each panel's own average offense and defense, so every team is read against its conference.

p = (
    ggplot(ratings, aes("ortg", "drtg"))
    + geom_mean_lines(aes(x0="ortg", y0="drtg"), color="#9a9a9a", size=0.6)
    + geom_sdv_logos(aes(team="team_abbreviation"), league="nba", season=SEASON, height=0.08)
    + facet_wrap("conference")
)
p

Better, but the best defenses are at the bottom, logos at the edges are clipped (Denver, Brooklyn, Washington), and the axis titles are column names.

4. Point the axes the right way and add net-rating guides

scale_y_reverse puts the best defenses on top, so up and to the right is good on both axes, and expand on both scales leaves room for the edge logos. Net rating is the gap between the two numbers, so teams with the same net rating sit on a diagonal: faint lines at -10, -5, 0, +5 and +10 show it without a third axis. Each line is clipped to the data's range and drawn as data (geom_line), so the reversed scale moves it with the logos.

x_lo, x_hi = ratings["ortg"].min() - 1, ratings["ortg"].max() + 1
y_lo, y_hi = ratings["drtg"].min() - 1, ratings["drtg"].max() + 1
# the line drtg = ortg - net, cut to the box [x_lo, x_hi] x [y_lo, y_hi]
ends = (
    pl.DataFrame({"net": [-10, -5, 0, 5, 10]})
    .with_columns(
        x0=pl.max_horizontal(pl.lit(x_lo), pl.lit(y_lo) + pl.col("net")),
        x1=pl.min_horizontal(pl.lit(x_hi), pl.lit(y_hi) + pl.col("net")),
    )
    .filter(pl.col("x0") < pl.col("x1"))
)
guides = (
    ends.unpivot(["x0", "x1"], index="net", value_name="ortg")
    .with_columns(drtg=pl.col("ortg") - pl.col("net"))
    .join(pl.DataFrame({"conference": ["East", "West"]}), how="cross")  # the same guides in both panels
)
guide_labels = (
    ends.with_columns(ortg=pl.col("x0") + 0.9 * (pl.col("x1") - pl.col("x0")))  # near the lower-right end
    .with_columns(
        drtg=pl.col("ortg") - pl.col("net"),
        label=pl.when(pl.col("net") > 0).then(pl.format("+{}", "net")).otherwise(pl.col("net").cast(pl.Utf8)),
    )
    .join(pl.DataFrame({"conference": ["East", "West"]}), how="cross")
)

p = (
    ggplot(ratings, aes("ortg", "drtg"))
    + geom_line(aes(group="net"), data=guides, color="#d9d9d9", size=0.5)
    + geom_label(aes(label="label"), data=guide_labels, color="#a6a6a6", size=7, fill="white", label_size=0)
    + geom_mean_lines(aes(x0="ortg", y0="drtg"), color="#9a9a9a", size=0.6)
    + geom_sdv_logos(aes(team="team_abbreviation"), league="nba", season=SEASON, height=0.08)
    + facet_wrap("conference")
    + scale_x_continuous(expand=(0.05, 0))
    + scale_y_reverse(expand=(0.07, 0))
    + labs(x="Offensive rating (points per 100 possessions)", y="Defensive rating (allowed per 100)")
)
p

5. Polish and tell the story

A theme strips the chart junk (minor grid, panel background) and sets the type; the strip titles become plain bold labels. The title states the finding, the subtitle explains the reading, and the caption carries the source. The two conference leaders get their net rating printed under the logo.

leaders = (
    ratings.group_by("conference")
    .agg(pl.all().sort_by("net").last())
    .sort("conference")
    .with_columns(
        note=pl.format("{} +{}", "team_abbreviation", pl.col("net").round(1))  # both leaders are above zero
    )
)
top = leaders.sort("net", descending=True).row(0, named=True)
east = leaders.filter(pl.col("conference") == "East").row(0, named=True)
title = f"{top['team_abbreviation']} stood alone at +{top['net']:.1f}; {east['team_abbreviation']} led a crowded East"

blog_theme = theme_minimal(base_size=10) + theme(
    figure_size=(9, 5.4),
    plot_title=element_text(weight="bold", size=14),
    plot_subtitle=element_text(color="#6b6b6b", size=9.5),
    plot_caption=element_text(color="#6b6b6b", size=7.5),
    strip_text=element_text(weight="bold", size=11, ha="left"),
    axis_title=element_text(color="#6b6b6b", size=9),
    axis_text=element_text(color="#8a8a8a"),
    panel_grid_minor=element_blank(),
    panel_grid_major=element_line(color="#efefef"),
    plot_title_position="plot",
)
chart = (
    p
    + geom_text(aes(label="note"), data=leaders, size=8, fontweight="bold", ha="left", nudge_x=0.75, color="#333333")
    + labs(
        title=title,
        subtitle=f"Points scored and allowed per 100 possessions, {LABEL} regular season; up and right is better.\n"
        "Dashed lines: each conference's average. Diagonals: net rating, in steps of 5.",
        caption="Data: hoopR (ESPN) via sportsdataverse-py  |  Chart: sdvplot + plotnine",
    )
    + blog_theme
)
chart

6. Export for the blog and for Instagram

ggplot.save takes the size in inches and a dpi: 8 x 4.5 in at 200 dpi is the blog's 1600 x 900. The square post re-lays the same plot with one more layer: facet_wrap(..., ncol=1) stacks the conferences, and a theme tweak sets the square figure size.

blog = OUT / "nba_net_rating_1600x900.png"
chart.save(blog, width=8, height=4.5, dpi=200, verbose=False)

square = OUT / "nba_net_rating_1080x1080.png"
(chart + facet_wrap("conference", ncol=1) + theme(figure_size=(7.2, 7.2))).save(
    square, width=7.2, height=7.2, dpi=150, verbose=False
)
for f in (blog, square):
    print(f.name, PILImage.open(f).size)
Image(blog, width=800)

nba_net_rating_1600x900.png (1600, 900)
nba_net_rating_1080x1080.png (1080, 1080)

The square cut, stacked:

Image(square, width=540)

Run it yourself

Download the notebook (outputs cleared) or open it on GitHub.